Predicting post-synaptic activity in proteins - ResearchGate

BIOINFORMATICS
Databases
Vol. 21 Suppl. 2 2005, pages ii19–ii25
doi:10.1093/bioinformatics/bti1102
Predicting post-synaptic activity in proteins with data mining
Gisele L. Pappa 1 , Anthony J. Baines 2 and Alex A. Freitas 1,∗
1 Computing Laboratory, University of Kent, Canterbury CT2 7NF, UK and 2 Department of Biosciences,
University of Kent, Canterbury CT2 7NJ, UK
ABSTRACT
Summary: The bioinformatics problem being addressed in this paper
is to predict whether or not a protein has post-synaptic activity. This
problem is of great intrinsic interest because proteins with postsynaptic
activities are connected with functioning of the nervous
system. Indeed, many proteins having post-synaptic activity have
been functionally characterized by biochemical, immunological and
proteomic exercises. They represent a wide variety of proteins with
functions in extracellular signal reception and propagation through
intracellular apparatuses, cell adhesion molecules and scaffolding proteins
that link them in a web. The challenge is to automatically discover
features of the primary sequences of proteins that typically occur in
proteins with post-synaptic activity but rarely (or never) occur in proteins
without post-synaptic activity, and vice-versa. In this context, we
used data mining to automatically discover classification rules that predict
whether or not a protein has post-synaptic activity. The discovered
rules were analysed with respect to their predictive accuracy (generalization
ability) and with respect to their interestingness to biologists
(in the sense of representing novel, unexpected knowledge).
Contact: A.A.Freitas@kent.ac.uk
1 INTRODUCTION
One of the great challenges of our era is to predict the functions of
proteins based on their primary sequence. This is a very difficult
problem, since the relationship between protein sequence and function
is very complex (Gerlt and Babbitt, 2000; Devos and Valencia,
2000; Nagl, 2003). Indeed, although there is a vast amount of data
stored in protein databases, there is still a large gap between the huge
amount of data about protein sequences and the knowledge necessary
for understanding the process of protein folding and associated
protein functions.
Intuitively, however, protein databases contain important ‘hidden
relationships’ (knowledge) between protein sequence and protein
function. There is a clear and urgent motivation for discovering this
hidden knowledge from protein databases for a number of reasons,
such as a better understanding of diseases, designing more effective
medical drugs, etc. This creates both a need and an opportunity
to apply data mining techniques to the problem of automatically
discovering knowledge from protein databases.
Data mining is a multi-disciplinary field, which consists of using
methods of several research areas (arguably, mainly machine learning
and statistical pattern recognition) to extract interesting knowledge
from real-world datasets (Witten and Frank, 2000; Fayyad et al.,
1996).
∗ To whom correspondence should be addressed.
This paper proposes a data mining approach to the problem of
predicting whether or not a protein has post-synaptic activity, based
on features of the protein’s primary sequence. The proposed approach
will be described later, in Sections 3 and 4. In this Introduction
we only emphasize a major difference between the proposed data
mining approach and a more traditional bioinformatics approach for
predicting protein function, as follows.
In general, the approach most used to predict the function of a new
protein—for which we know only its sequence—consists of performing
a similarity search in a protein database. In essence, the program
finds the most similar protein(s) to the new protein, and if that similarity
is higher than a threshold, the function of the most similar
protein is transferred to the new protein. Although this approach is
very useful in many cases, it also has some limitations, as follows.
First, it is well-known that two proteins might have very similar
sequences and perform different functions, or have very different
sequences and perform the same or similar function (Syed and Yona,
2003; Gerlt and Babbitt, 2000). Second, the proteins being compared
may be similar in regions of the sequence that are not determinants
of function (Schug et al., 2002). Third, the prediction of function is
based only on sequence similarity, ignoring many relevant biochemical
properties of proteins (Karwath and King, 2002; Syed and Yona,
2003). Fourth, it does not produce a model for predicting function,
and so it does not give insights into the relationship between the
sequence, biochemical properties and function of proteins.
Another approach consists of inducing, from protein data, a
model describing (in a very summarized form) the data, so that
new proteins can be classified by the model. This is the data
mining approach followed in this project. We emphasize that
this model-induction approach aims at complementing—rather than
replacing—the conventional similarity-based approach. In any case,
the model-induction approach followed in this project has two
important advantages (King et al., 2001). First, it can predict the
function of a new protein even in the absence of sequence similarity
between that protein and other proteins with known function. Second,
if the discovered model is expressed in a comprehensible form (which
is the case in this research, where knowledge is expressed by intuitively
comprehensible IF-THEN rules), it can be used by biologists,
to give new insights and possibly suggest new biological experiments.
Indeed, in this research the rules discovered by a data mining
algorithm were not only automatically evaluated with respect to their
predictive accuracy—as is usual in data mining—but also manually
interpreted in the context of relevant biochemical knowledge, in order
to determine how interesting they were with respect to providing
novel insights—unknown in the current literature—about the relationship
between some protein sequence patterns and post-synaptic
activity. This two-criteria evaluation reinforces the inter-disciplinary
© The Author 2005. Published by Oxford University Press. All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org ii19

G.L.Pappa et al.
Fig. 1. Main elements involved in pre-synaptic and post-synaptic activity.
A synapse is the point where two nerve cells communicate with each other by
transmission of a chemical known as a neurotransmitter. The main elements
found in synapses are shown in Figure 1. The cells are held together by
cell adhesion molecules (1). In the cell where the signal is coming from
(the pre-synaptic cell) neurotransmitters are stored in bags called synaptic
vesicles. When signals are to be transmitted from the pre-synaptic cell to the
post-synaptic cell, synaptic vesicles fuse with the pre-synaptic membrane and
release their contents into the synaptic cleft between the cells. The transmitters
then diffuse within the cleft, and some of them meet a post-synaptic receptor
(2), which recognises them as a signal. This activates the receptor, which
then transmits the signal on to other signalling components such as voltagegated
ion channels (3), protein kinases (4) and phosphatases (5). To ensure
that the signal is terminated and to clear up residual neurotransmitters after
the signal has terminated, transporters (6) remove neurotransmitters from the
cleft. Within the post-synaptic cell, the signalling apparatus is organised by
various scaffolding proteins (7).
nature of this project, which is, of course, a desirable feature in a
bioinformatics project.
The remainder of this paper is organised as follows: Section 2
describes the target biological problem; Section 3 describes the
preparation of the dataset for mining purposes, by using protein
data available in UniProt/SwissProt and Prosite; Section 4 discusses
the proposed data mining approach and the corresponding analysis
of results; finally, Section 5 reports the conclusions and future
research directions.
2 THE TARGET BIOLOGICAL PROBLEM
In essence, post-synaptic sites represent points where one nerve cell
receives signals from another. As indicated in Figure 1, multiple
types of proteins are expected to be found at these sites for reception
and propagation of signals, and for joining the two nerve cells to each
other. Note that Figure 1 is a very minimal summary of the types of
proteins found in postsynaptic sites.
The bioinformatics problem being addressed in this paper is to predict
whether or not a protein has post-synaptic activity. This problem
is of great intrinsic interest because proteins with post-synaptic activities
are connected with functioning of the nervous system. Indeed,
many proteins having post-synaptic activity have been functionally
characterized by biochemical, immunological and proteomic exercises
[see e.g. Husi et al. (2000) and Walikonis et al. (2000)], and are
ii20
now extensively catalogued and annotated in the Uniprot/SwissProt
database. They represent a wide variety of proteins with functions in
extracellular signal reception and propagation through intracellular
apparatuses, cell adhesion molecules and scaffolding proteins that
link them in a web.
The challenge is to automatically discover features of proteins’
primary sequences that typically occur in proteins with post-synaptic
activity but rarely (or never) occur in proteins without post-synaptic
activity, and vice-versa. These discovered features constitute the
essence of the knowledge discovered by a data mining algorithm.
If the algorithm is successful in discovering knowledge with a high
predictive power, that knowledge can be used to accurately discriminate
between the two classes of proteins. In addition, and
most important, the fact that in this project discovered knowledge
will be expressed in a comprehensible form—as mentioned in the
Introduction—represents potentially valuable knowledge by itself,
because such knowledge can potentially give new insights to biologists
about which sequence features are predictive of post-synaptic
activity.
3 METHODS
The goal of this project is to predict—with a data mining algorithm—whether
or not a protein has post-synaptic activity. The algorithm is used to discover
interesting relationships between sequence features that are often present in
post-synaptic proteins but usually absent in proteins without post-synaptic
proteins, and vice-versa. Therefore, in order to discover this kind of knowledge,
we need not only a set of post-synaptic proteins but also a control set
of proteins which do not have post-synaptic activity.
In data mining terminology, the set of proteins with post-synaptic activity
is called the set of positive examples, whereas the control set of proteins
(without post-synaptic activity) is called the set of negative examples. The
problem of finding sequence features that discriminate between these two
kinds of proteins is then cast as a classification problem, where the goal is
to predict the value of a class attribute for each example (protein) based on
a set of predictor attributes for that example. The classes are whether or not
a protein has post-synaptic activity, and the predictor attributes are mainly
Prosite patterns, as described below.
More precisely, the dataset mined in this project was constructed in
two phases. In the first phase we carefully selected relevant proteins
from the UniProt database. UniProt (Universal Protein Resource) was
created by the union of Swiss-Prot, TrEMBL and PIR databases, and
is a repository for protein sequence and functional data (Uniprot, 2004,
http://www.ebi.uniprot.org/index.shtml). UniProt was chosen as the source
of the data to be mined because it is the world standard non-redundant,
comprehensive protein sequence database. UniProt is divided in three
database layers: UniParc (UniProt Archive), UniProt Knowledgebase (Uni-
Prot/SwissProt) and UniProt NREF (UniRef). In this research we have used
UniProt/SwissProt, which is the richest annotated layer. One of the great
advantages of this layer is the comprehensive annotation of SwissProt. This
is curated to ensure minimal redundancy, accurate and comprehensive annotation
of function, expression, sequence features (e.g. domain structures),
literature references, links to other databases, etc.
The second phase of data preparation used the links from Unit-
Prot/SwissProt to the Prosite database, in order to retrieve information from
Prosite that was used to create the predictor attributes. These two phases are
described in detail next.
3.1 Phase 1: selecting positive and negative examples
Table 1 shows the queries submitted to UniProt/SwissProt in order to select the
set of positive examples (proteins with post-synaptic activity) and the set of
negative examples (proteins without post-synaptic activity). For each query,
the table shows the specification of the query and the number of examples

Table 1. Queries submitted to UniProt/SwissProt to create the dataset
Query No. Query specification # Examples
Selection of positive examples
1 Post-synaptic !toxin 356
Total number of positive examples 356
Query No. Query/Keywords # Examples
Selection of negative examples
2 Heart !(result of query 1) 3106
3 Cardiac !(results of queries 1, 2) 331
4 Liver !(results of queries 1, 2, 3) 2794
5 Hepatic !(results of queries 1, 2, 3, 4) 256
6 Kidney !(results of query 1, 2, 3, 4, 5) 988
Total number of negative examples 7475
(proteins) returned by that query. The specification of each query consists of
keywords and the logical operator ‘NOT’ (!).
Note that the queries did not specify any specific species, i.e. proteins from
all species were considered. This was done in order to maximize the number
of examples in the data to be mined. Positive examples were selected not
only by the presence of the keyword ‘post-synaptic’ but also by the absence
of the keyword ‘toxin’. The reason for this latter criterion is that several
entries in UniProt/SwissProt refer to the toxin α-latrotoxin. This protein acts
post-synaptically, but it is not, of course, a post-synaptic protein.
The selection of the negative examples was more difficult, since of course
the UniProt/SwissProt entries do not have an explicit ‘not post-synaptic’
keyword. The trivial ‘solution’ of retrieving all entries that do not explicitly
have the keyword ‘post-synaptic’ is not satisfactory, for two reasons.
First, this would produce a very large number of negative examples, which
would be much larger than the number of positive examples. This would create
a dataset with an extremely unbalanced class distribution, and it would be
very difficult for the classification algorithm to discover rules that correctly
classify the positive (post-synaptic) class. Second, and most important, many
of the negative examples would be ‘trivial’, leading to uninteresting, trivial
rules for the discrimination between positive and negative examples. For
instance, there is no need to discover rules discriminating plant proteins from
post-synaptic proteins, since it is obvious that plants do not have a nervous
system. Hence, including plant proteins in the set of negative examples would
only contribute to the discovery of uninteresting, trivial classification rules.
The goal is to select a set of negative examples where, although the proteins
do not have post-synaptic activity, they have some characteristics that
could be confused with some of the characteristics of post-synaptic proteins
(the positive examples), making it difficult to discriminate between these two
kinds of proteins. Intuitively, the higher the similarity between positive and
negative examples, the harder the classification problem, and so the stronger
the motivation to use a data mining algorithm to discover interesting classification
rules representing novel knowledge to a biologist. (Recall that the
ultimate goal is to automatically discover interesting, novel rules that provide
new insight to biologists about which sequence features are most correlated
with the presence or absence of post-synaptic activity.)
Note that, although the positive and negative examples must have enough
similarity to lead to the discovery of interesting, non-trivial rules, they also
must be different enough to allow the discovery of reliable classification rules
for each class. Hence, the challenge is to find a good trade-off between these
two goals.
In this spirit, the negative examples were selected by using the keywords
‘heart’, ‘cardiac’, ‘liver’, ‘hepatic’ and ‘kidney’. Consider, for instance, the
query 4 in Table 1: ‘Liver !(results of queries 1, 2, 3)’. This query means
Predicting post-synaptic activity in proteins with data mining
that all UniProt/ SwissProt entries that had the keyword ‘liver’ and were not
already included in the results of the previous queries (1, 2, 3) were selected
as negative examples. The latter selection criterion was necessary to avoid the
possibility that two or more copies of the same data entry—i.e. a data entry
having two or more of the above-mentioned keywords—were duplicated in
the set of negative examples.
The previously-mentioned five keywords were chosen as the basis for
obtaining negative examples for two reasons. First, it is known that, in general,
proteins found in these sites do not present post-synaptic activity. Second,
proteins found in these sites often have some characteristics similar to postsynaptic
proteins. Indeed, in the context of this project, many of the same types
of proteins represented at post-synaptic sites (kinases/phosphatases/channels,
etc.) are present in abundance in heart, liver and kidney tissues.
It should be noted that the queries listed in Table 1 searched for the corresponding
keywords in all the fields of the UniProt/ SwissProt entries. This
means that the selection criteria are not perfect, since those keywords could
be present in fields where the presence of the keyword does not mean that the
protein has the corresponding function/characteristic. This potentially introduces
some ‘noise’ in the dataset being mined. However, this was necessary,
because queries searching for keywords only in the field ‘KEYWORD’ of
UniProt/ SwissProt returned too few proteins. In any case, the amount of
noise introduced by the imperfect selection criteria seems to be relatively
small, since the data mining algorithm was able to discovery quite accurate
classification rules, as will be shown later.
3.2 Phase 2: generating the predictor attributes
Once the set of examples to be mined has been selected from Uniprot/Swissprot,
the next step was to generate a set of predictor attributes
representing relevant properties of the sequences those proteins. The predictor
attributes must have a good predictive power and facilitate easy interpretation
by biologists. In this project we have focused mainly on generating attributes
based on Prosite patterns associated with the proteins—a type of attribute
satisfying both the previously-mentioned properties. The Prosite database
stores significant patterns and profiles that help to identify the family of a
new protein (Hulo et al., 2004). We decided to use attributes based only on
Prosite patterns, and not Prosite profiles, for two reasons. First, the matching
between a Prosite pattern and a protein can be exactly computed, producing
a simple binary attribute—i.e. the pattern either occurs or does not occur in
a given protein. This reduces the size of the search space for the data mining
algorithm and simplifies the interpretation of the rule by biologists. By
contrast, the matching between a Prosite profile and a protein is an approximate
matching, and the data mining algorithm would have to search in a
correspondingly much larger search space. Second, the use of both patterns
and profiles would lead to a very large number of attributes, which would
again expand the size of the search space, and so would significantly increase
the risk of overfitting the induced model to the data. It should be noted that,
even considering only the binary attributes derived from Prosite patterns, this
led to 443 attributes (as explained next), which corresponds to a huge search
space of size 2443 .
For each protein selected in phase 1 (regardless of the protein being a
positive or a negative example), we retrieved all Prosite entry id’s that occurred
in the field database cross-references (DR) of UniProt/SwissProt. For each
Prosite entry id, we ‘followed the link’ from UniProt/SwissProt to Prosite, in
order to access information about that Prosite entry. Once this was done for
all proteins selected in phase 1, we had a large set of Prosite entries. We then
selected, for use as predictor attributes, the entries that:
(i) were marked as a ‘pattern’ in the ID line of that entry in the Prosite
database;
(ii) were not commented (in the CC line of the Prosite database) as
very general patterns (/SKIP_FLAG = TRUE)—it was necessary to
exclude those patterns because they appear in almost all proteins, and
so are not useful to discriminate between the two classes of proteins;
(iii) occurred in at least two proteins of the dataset being mined—this
was necessary to remove extremely specific patterns, occurring in
ii21

G.L.Pappa et al.
just one protein of the dataset, which do not have any generalisation
power.
Finally, each of the selected Prosite patterns was encoded as one binary
attribute of the dataset being mined, taking on the value ‘yes’ or
‘no’ for each protein—indicating whether or not the pattern occurs in that
protein, respectively.
Note that, a few proteins in the dataset to be mined did not have any Prosite
pattern, i.e. they had the value ‘no’ for all predictor attributes based on Prosite
patterns. These proteins were removed from the data to be mined.
In addition to the previously-described attributes based on Prosite patterns,
we added to the dataset two simple predictor attributes derived directly from
the proteins’ sequences, namely the sequence length and the molecular weight
of the protein (both attributes are available from the corresponding fields in
UniProt/SwissProt entries). Other kinds of attributes will be considered in
future research, but for now it is interesting to note that even the current set of
predictor attributes is enough to discover quite accurate classification rules,
as will be shown later.
After all this data preparation process, we ended up with a dataset composed
by 4303 examples (260 belonging to the positive class and 4043
belonging to the negative class) and 445 predictor attributes. These 445
attributes include 443 Prosite patterns, the sequence length and the molecular
weight of each protein.
4 RESULTS
In order to discover knowledge from the dataset described in the
previous section, we have used the well-known C4.5Rules rule induction
algorithm (Quinlan, 1993). This algorithm was chosen as the
data mining algorithm in our experiments mainly because it produces
comprehensible knowledge, represented by a set of high-level,
easily-interpretable classification rules of the form: IF (conditions)
THEN (class). This kind of rule has the intuitive meaning that, if an
example (protein) satisfies the conditions in the rule antecedent, the
example is assigned to the class predicted by the rule consequent.
It should be noted that the comprehensibility of discovered knowledge
is a very important issue in bioinformatics [see e.g. Mirkin
and Ritter (2000), Clare and King (2002) and Sebban et al.
(2002)], because the discovered knowledge should be interpreted
and validated by biologists, rather than being blindly trusted as a
‘black box’.
We used the default parameters of C4.5Rules. The classification
rules discovered by C4.5Rules were evaluated according to two criteria,
namely predictive accuracy and interestingness to biologists, as
follows. Predictive accuracy was estimated by a well-known 10-fold
cross-validation procedure (Witten and Frank, 2000), as usual in data
mining. In essence, the dataset was divided into 10 partitions, with
approximately the same number of examples (proteins) in each partition.
In the i-th iteration, i = 1, 2, ..., 10, the i-th partition was used
as the test set and the other 9 partitions were temporarily merged and
used as the training set. In each iteration C4.5Rules discovered a rule
set from the training set and used that rule set to classify examples in
the test set (unseen during training), in order to evaluate the generalisation
ability of discovered knowledge. The classification accuracy
rate of the discovered rules can then be computed as the average
accuracy rate over the 10 test sets, and this is the measure of predictive
accuracy most popular in the literature. In our experiments, this
produced an accuracy rate of 97.85%.
It should be noted, however, that in the context of this project this
traditional measure of accurate rate is not a very effective one. The
reason is that the class distribution is very unbalanced: only 6.4% of
the examples have the positive class. Hence, as a baseline solution for
ii22
this classification problem, the ‘majority classifier’—which predicts
the majority (negative) class for all examples—would trivially obtain
an accuracy rate of 93.9%, without providing any insight about the
relationship between the predictor attributes and the classes.
Therefore, we use a more ‘demanding’ measure of predictive
accuracy, for which a high value can be obtained only by accurately
classifying examples of both classes. The measure in question is the
product: true positive rate (TPR) × true negative rate (TNR) (Hand,
1997). These terms (which are sometimes referred to as Sensitivity
and Specificity, respectively) are defined as follows.
TPR = TP/(TP + FN) TNR = TN/(TN + FP),
where
TP = number of true positives—i.e. the number of examples that
were predicted as positive class by the discovered rule set,
and indeed have the positive class;
FN = number of false negatives—i.e. the number of examples that
were predicted as negative class, but actually have the positive
class;
TN = number of true negatives—i.e. the number of examples that
were predicted as negative class, and indeed have the negative
class;
FP = number of false positives—i.e. the number of examples that
were predicted as positive class, but actually have the negative
class.
In our experiments the average values (over the 10 iterations of the
cross-validation procedure) of the TPR and TNR were 0.85 and 0.98
respectively, resulting in the final measure of predictive accuracy as
TPR × TNR = 0.84 (with a standard deviation of 0.09). Note that,
the baseline majority classifier obtains TPR × TNR = 0 × 93.9 = 0,
i.e. it is very strongly penalized (as it should be) for never predicting
the positive class.
It should also be noted that, although the vast majority of the data
mining literature focuses on measuring only the predictive accuracy
of the discovered rules, the ultimate goal of data mining is to discover
knowledge that is comprehensible and interesting (novel, unexpected)
to the user (Fayyad et al., 1996; Han and Kamber, 2001). We
emphasize that a very accurate rule will not be useful to the user if
it represents a previously known pattern. Consider, for instance, the
following hypothetical example. In a hospital’s medical database a
data mining algorithm could discover the rule: IF (patient is pregnant)
THEN (patient’s gender is female). This rule is extremely accurate,
but it is also completely useless, since it represents an obvious pattern.
As a real-world example of the difficult of discovering novel,
unexpected rules, (Tsumoto, 2000) reports that, in experiments with
two medical datasets,

Table 2. Rules discovered by C4.5Rules
Id Classification rule
32 IF (NEUROTR_ION_CHANNEL = yes) THEN (class = yes)
19 IF (CADHERIN_1 = yes) AND (920 < seq_length 699)
THEN (class = yes)
10 IF (G_PROTEIN_RECEP_F1_1 = yes)
AND (11287 < mol_weigth 318)
THEN (class= yes)
21 IF (G_PROTEIN_RECEP_F3_1 = yes)
AND (mol_weight

G.L.Pappa et al.
31.4%. The low accuracy of these two rules comes from the fact that
ser/thr phosphatases and G-protein coupled receptors are expressed
in every human cell, and not just post-synaptically.
The unexpected rules are much more complicated, but they
are very surprisingly accurate. Therefore, in general they represent
interesting knowledge to biologists who are experts in
post-synaptic proteins.
For example, Rule 7 states that if 8 specific Prosite signatures are
absent, then the protein is not post-synaptic with 99.8% accuracy.
Similarly, Rule 2 states the same thing with 12 Prosite signatures.
These rules could not have been predicted a priori with just biological
knowledge. What Rules 2 and 7 do is to take a number of
Prosite signatures that appear in individual ‘expected’ rules and combine
them in a way that says that, when none of those signatures is
present, then the proteins are not post-synaptic. This has real utility
in classifying novel proteins, excluding them from the post-synaptic
class.
In order to better understand those rules, we also retrieved, from
the dataset being mined, the proteins which are exceptions to each
of those rules—i.e. proteins that satisfy the conditions in the rule
antecedent but have a class different from the one predicted by the
rule consequent. These exceptions are quite revealing.
An exception to Rule 7 is the SPOCK protein (Uniprot:
TIC1_MOUSE). This is an extracellular matrix (ECM) protein that
is associated with the post-synaptic area of pyramidal neurons. The
signatures in Rule 7 are membrane-associated or cytoplasmic; thus
they do not cover this ECM protein. The synaptic cleft is rather poor
in ECM proteins, so most other ECM proteins (collagen etc.) are
accurately included in this rule.
Rule 2 has some interesting exceptions, and again points to limitations
in the method used to generate the predictor attributes. The
protein b-Raf is a protein kinase that is ubiquitous in animal tissues,
and has a role in mitogenic signalling. It is also found in synaptic
structures. In neurons, it is thought to be part of the system that
responds to growth factors such as NGF. In this sense it is not classically
part of the system that responds to neurotransmitters, but rather
has a role in development and maintenance of the nervous system.
B-raf from human and mouse (Uniprot entries: BRAF_HUMAN and
BRAF_MOUSE) are exceptions to Rule 2. This exception reflects
both the ubiquity of b-Raf and the fact that it represents the nature
of the signalling pathways it is involved in.
5 CONCLUSIONS
This paper proposed a data mining approach to generate comprehensible
rules that predict whether or not a protein has post-synaptic
activity, based on Prosite patterns occurring (or not) in the protein,
as well as on a couple of simple protein properties computed directly
from the protein’s primary sequence (namely the sequence length
and the molecular weight of the protein).
The discovered rules were evaluated with respect to both their predictive
accuracy and their degree of surprisingness (unexpectedness)
to the user. The discovered rules were very successful with respect to
predictive accuracy. The main contribution of this paper, however, is
the analysis of the rules with respect to their surprisingess. Although
this is a very important issue in data mining (as discussed earlier),
and particularly crucial in the context of scientific discovery, this
issue is largely ignored in virtually all the literature about prediction
of protein function from sequence.
ii24
From a biological perspective, the discovered rules overall reveal
interesting features of this approach to mining functional data from
Uniprot/SwissProt. A number of expected rules accurately predict
some aspects of post-synaptic function. Other rules (unexpectedly)
can exclude post-synaptic function with astonishing accuracy. Still
other rules indicate the limitation of this approach. The lack of
voltage-gated ion channel Prosite patterns (related to type 3 proteins
in the notation of Fig. 1) reflects limitations in Prosite:
future approaches to this problem will need to consider this. In
the future we also plan to generate a more diverse set of predictor
attributes, capturing information about other relevant properties of
protein sequences.
A direction for future research would be to estimate the ‘interestingness’
of the discovered rules by using some data-driven rule
interestingness measures proposed in the literature. Then we would
be able to automatically rank the discovered rules according to
those interestingness measures, and present the rules to the user
in decreasing order of estimated interestingness. We could also
measure the correlation between the value of those data-driven interestingness
measures and the subjective, real interest of the rules to
a biologist. This would allow us to evaluate how effective those
data-driven interestingness measures are in the sense of being good
estimators of the real human interest in the rules. It would also
be interesting to analyse the rules discovered by other data mining
algorithms.
ACKNOWLEDGEMENTS
G.L.P. is financially supported by CAPES (a Brazilian researchsupport
agency), process number 1650-02-5.
Conflict of Interest: none declared.
REFERENCES
Clare,A. and King,R.D. (2002) Machine learning of functional class from phenotype
data. Bioinformatics, 18, 160–166.
Devos,D. and Valencia,A. (2000) Practical limits of functional prediction. Proteins, 41,
98–107.
Fayyad,U.M., Piatetsky-Shapiro,G. and Smyth,P. (1996) From data mining to
knowledge discovery: an overview. In: Fayyad,U.M., Piatetsky-Shapiro,G. and
Smyth,P. (eds), Advances in Knowledge Discovery and Data Mining. AAAI/MIT,
pp. 1–34.
Gerlt,J.A. and Babbitt,P.C. (2000) Can sequence determine function? Genome Biol.,
1(5), reviews 0005.1–reviews 0005.10.
Han,J. and Kamber,M. (2001) Data Mining: Concepts and Techniques. Morgan
Kaufmann.
Hand,D.J. (1997) Construction and Assessment of Classification Rules. Wiley.
Hulo,N. et al. (2004) Recent improvements to the PROSITE database. Nucleic Acids
Res., 32, D134–D137.
Husi,H. et al. (2000) Proteomic analysis of NMDA receptor–adhesion protein signaling
complexes. Nat. Neurosci., 3, 661–669.
Karwath,A. and King,R.D. (2002) Homology induction: the use of machine learning to
improve sequence similarity searches. BMC Bioinformatics, 3, 11.
King,R.D. et al. (2001) The utility of different representations of protein sequence for
predicting functional class. Bioinformatics, 17, 445–454.
Mirkin,B. and Ritter,O. (2000) A feature-based approach to discrimination and prediction
of protein folding groups. In: Suhai,S. (ed.), Genomics and Proteomics:
Functional and Computational Aspects. Kluwer/Plenum, pp. 157–177.
Nagl,S. (2003) Function prediction from protein sequence. In Orengo,C.A., Jones,D.T.
and Thornton,J.M. (eds), Bioinformatics: Genes, Proteins & Computers, Bios, pp.
65–79.
Quinlan,J.R. (1993) C4.5: Machine Learning Programs. Morgan Kaufmann.
Schug,J. et al. (2002) Predicting gene ontology functions from ProDom and CDD protein
domains. Genome Res., 12, 648–655.

Sebban,M. et al. (2002) A data mining approach to spacer oligonucleotide typing of
Mycobacterium tuberculosis. Bioinformatics, 18, 235–243.
Syed,U. and Yona,G. (2003) Using a mixture of probabilistic decision trees for direct
prediction of protein function. In Proceedings of the 2003 Conference on Research in
Computational Molecular Biology (RECOMB-2003), Berlin, Germany, pp. 289–300.
Tsumoto,S. (2000) Clinical knowledge discovery in hospital information systems: two
case studies. In Proceedings of the Fourth European Conference on Principles
Predicting post-synaptic activity in proteins with data mining
and Practice of Knowledge Discovery in Databases (PKDD-2000), LNAI 1910,
652–656. Springer-Verlag.
Uniprot: The Universal Protein Resource. (2004)
Walikonis,I.R.S. et al. (2000) Identification of proteins in the postsynaptic density
fraction by mass spectrometry. J. Neurosci., 20, 4069–4080.
Witten,H. and Frank,E. (2000) Data Mining: Practical Machine Learning Tools and
Techniques. Morgan Kaufmann.
ii25